AgentSkin Suite

Semantic JSON Pruning

The core protocol. Recursively prunes JSON/HTML into high-density Markdown skins.

• Creator: Shawn Nichols Sr. (Nichols Transco LLC)
• Savings: 60-88% on rich APIs
• Auto-Classification: 11 URL rule families
• 3-Layer Config: Builtin → User → Project override semantics
• MCP Tools (4): fetch_optimized_data, skin_reasoning, classify_url, strip_ansi

Rule-Driven CLI Compaction

Strips ANSI codes, compacts dependency trees, truncates repeated patterns, preserves head/tail of large outputs.

• Creator: Vincent Koc (MIT License)
• Savings: Up to 99.97% on large outputs
• Rules: 136 rules, 143 fixtures
• Bash Hook: Transparent CLI optimization via opt prefix
• Pipeline: 3,030 fixtures/sec (0.33ms avg)
• MCP Tools (3): reduce, estimate_tokens, apply_json_semantic (merged into unified server)

Output Compression via Prompt Engineering

Makes agents talk like cavemen. Same answers, 65% fewer output tokens. Brain still big. Mouth small.

• Creator: Julius Brussee
• Savings: 65% average output token reduction
• Memory Compression: 46% smaller files, saved forever
• Skills: caveman, caveman-commit, caveman-review, caveman-stats, caveman-compress, cavecrew
• Agents: Claude Code, Codex, Gemini, Cursor, Windsurf, Cline, Copilot, 30+ others

Full Stack Setup

Three commands. Full token optimization stack.

# 1. Bash hook (transparent CLI optimization)
echo "source $(pwd)/.agents/hooks/bash-optimizer.sh" >> ~/.bashrc && source ~/.bashrc

# 2. Caveman skills — already active via AGENTS.md
# (auto-discovered by Claude Code, Codex, Cursor, etc.)

# 3. MCP server — add to agent config (unified 7-tool server)
{
  "mcpServers": {
    "agentskin-suite": {
      "command": "npx",
      "args": ["-y", "agentskin@latest"]
    }
  }
}
            

Expected Savings (Per Session)

• Bash Hook: 17.1% net (up to 99.97% on large outputs)
• Caveman: 65% of output tokens
• MCP (AgentSkin): 60-88% per API call
• Combined: 60-80% of wasted context reclaimed